Field Notes
Notes on building
an AI marketing OS.
Threads on attribution, ad creative, AI video generation, and other things we're chewing on while shipping Orbit.
The Agency Reporting Stack in 2026: What to Use and Why
A client report in 2026 has four kinds of columns: what the platforms claim, what the warehouse observed, what customers said, and what an experiment proved. Each comes from a different tool category, moves on a different cadence, and answers a different question. The categories, what belongs in weekly versus monthly, the one-number-per-source rule that keeps the report honest, and where an agent layer like Orbit fits without replacing any of it.
The Analyst Agent: How Orbit Ingests GA4, PostHog, Mixpanel, Stripe, and Your Database to Build a Full-Funnel View
How Orbit's analyst reads a brand's analytics: which source it asks for first, why conversion definitions come before revenue and revenue comes before product analytics, how it reconciles them into an attribution mix and a visitors-to-retained funnel with revenue from the payment processor, what it writes down as the brand's analytics map, and when it pulls a guide. Present tense for what ships in orbit-mcp today; designed-to for what the agent runtime adds.
Budget Pacing: How to Manage Ad Spend Across Platforms Without Overspending
Google can spend 2x your daily budget on any day and 30.4x in a month. Meta can spend 75% over on a day and 7x in a Sunday-to-Saturday week. TikTok paces a lifetime budget unevenly and asks you to cap edits at 30 to 40 percent. The documented overdelivery rules per platform, the pacing math that turns them into one monthly number, what a budget edit does to learning, and the alerts an agent should run every morning.
GA4 Attribution Models: Data-Driven vs Last Click and What the Difference Actually Means
GA4 has three models left, two lookback settings that most teams have never opened, and since January 2026 a per-key-event override that changes what Google Ads bids on. What data-driven attribution actually computes, which reports the model touches and which it does not, how the August 2026 custom lookback windows work, and how to read a GA4 attribution report next to a Google Ads report without concluding one of them is broken.
Meta Custom Audiences 2026: Pixel vs CAPI vs List Upload
Under Advantage+ audience, a custom audience you include is a suggestion and a custom audience you exclude is a wall. That one fact reorders every use of website, list, and engagement audiences. What each source feeds, how list uploads and value-based lookalikes actually work now, why exclusions are the highest-leverage setting left, and why all of it depends on event match quality.
Target ROAS vs Target CPA: When to Use Which (and When to Use Neither)
Google's own eligibility floors, the 14,584-account dataset showing that targets often cost you performance, and the August 17, 2026 change that makes a target a reservation price instead of an aspiration. When target CPA fits, when target ROAS fits, when maximize conversions or value without a target is the right call, and how to set a target from reconciled revenue rather than from last month's dashboard.
Attribution for Subscription Businesses: LTV vs CAC and the Payback Period Frame (2026)
The click happens once and the money arrives for months. Why first-payment ROAS misleads, why the 3:1 rule is a 2011 artifact, how to compute payback by cohort with gross margin and churn split out, and how to feed lifetime value back into Google's lifecycle goals and Meta's value signals so the platforms bid on it.
Creative Testing Under Andromeda: A System for Finding the 5% That Wins (2026)
About 5% of ad creatives become winners and they take 55% of spend. A testing system built for that math: concepts versus variations, spend floors before any read, a seven-day rule, a metrics ladder from hook to CPA, and why a 65% confidence 'winner' is usually noise.
Creator Sourcing, UGC, and Partnership Ads: How to Buy Creative That the Algorithms Want (2026)
Creator content is the cheapest high-hit-rate creative you can buy, and both Meta and TikTok now let you run it from the creator's handle with permission. How to source by content fit instead of follower count, what rights to buy, how partnership ads and Spark Ads work, what the FTC requires, and how to run the whole thing as a pipeline with a hit rate.
Demand Gen Campaigns 2026: Setup, Creative Specs, and What Changed When It Swallowed Video Action and Display
Demand Gen is now where YouTube, Shorts, Discover, Gmail, and the Display Network live in Google Ads. Video action campaigns were folded in during 2025, display follows in 2026, and lookalikes became a suggestion in March. The structure, the channel controls, the creative that works across surfaces, and how to read it against Meta.
Google Enhanced Conversions: What Actually Changed and How to Set It Up Right (2026)
Enhanced conversions for web and leads became one toggle in June 2026, multiple data paths can now run at once, and offline uploads are moving to Data Manager. What the diagnostics report is really telling you, why match rates collapse, and the setup that recovers the conversions smart bidding cannot see.
First-Party Data Attribution: Building Your Own Source of Truth After iOS 17 (2026)
Three platforms each claim the same order and Safari deletes the cookie in a week. The way out is a ledger you own: server-set first-party ids, every click id and UTM captured at landing, hashed identity at login and checkout, a warehouse join to platform cost, attribution rules you wrote, a survey for the channels with no click, and a holdout to calibrate it all.
GA4 + Google Ads: How to Import Conversions Properly in 2026
The link, the key-event import, and the native tag are three different things. Which one should bid, which one should observe, why the numbers never match, and how to align attribution, windows, and counting so they match closely enough to trust.
Google Ads Attribution Setup 2026: The Stack That Survives Cookies, iOS, and the API Deadlines
Google measurement is a stack now: the Google tag, conversion linker, enhanced conversions, consent mode v2, offline uploads through Data Manager, Customer Match, server-side tagging or the tag gateway, and data-driven attribution. What each layer covers, the 2026 API cutoffs that broke old integrations, the code for each path, and the weekly checks.
Incrementality Testing on a Small Budget: The Geo Holdout Method (2026)
How to measure what your ads actually cause without a six-figure vendor contract: pick the KPI from your source of truth, size the test with a power analysis, hold out 2 to 5 markets for at least one purchase cycle, and turn the result into a multiplier you apply to every platform-reported number.
Landing Pages for Paid Traffic: What the Page Has to Do Now That the Ad Picks the Audience (2026)
Under Andromeda the ad chooses who arrives, so the page has to keep the ad's promise in the first screen. Speed data from 30 million sessions, conversion benchmarks from 41,000 pages, Google's landing page rules, the URL controls most accounts skip, and how to test without fooling yourself.
The Algorithms Behind Meta Ads: Andromeda, GEM, Lattice, and How They Actually Fit Together (2026)
Meta ads is not one algorithm. It is a retrieval engine (Andromeda), a foundation model (GEM) that teaches hundreds of ranking models (Lattice, ARM), and an auction on top. Where each one sits, what each one reads from your account, and how to diagnose performance by stage.
Meta Advantage+ Sales Campaigns: When to Use Them, When to Run Manual, and What the Incrementality Data Says (2026)
Advantage+ sales campaigns take most Meta ecommerce spend and report a flattering ROAS. Haus's 640 geo tests say they win early, fade by the end of the test, and over-report against manual by 12 points. The split that works: automated harvesting with a budget ceiling, manual prospecting to feed it, and a holdout to price both.
Meta's Andromeda Update: What Actually Changed and What to Do About It (2026)
Andromeda rebuilt Meta's ad retrieval engine and turned creative into targeting. What broke in existing accounts, the data behind the ROAS drop, what changed in 2026, and the structure that works now.
Meta CAPI Setup 2026: Server-Side Events, Deduplication, and the EMQ Score
The pixel plus conversions api stack that meta's ranking models actually train on: which parameters move event match quality, how deduplication really works, the server code for a purchase from a payment webhook, what changed in 2025 and 2026, and the checks an agent runs weekly.
New Client Ad Audit Checklist: 42 Things to Check in the First Week (2026)
The first-week audit for any ad account you inherit: access and context, tracking on every platform, money and pacing, structure and targeting, creative, landing pages, and the baseline you report against. Every item has a threshold, so an agent can run it.
Performance Max 2026: The Campaign Structure That Actually Works
Performance Max got the controls it was missing: campaign-level negatives, brand exclusions, search terms, channel reporting, customer-list exclusions. Most accounts still run the 2023 setup. What the data says about brand cannibalization, conversion volume, and asset groups, and the structure to run now.
TikTok Events API Setup 2026: Deduplication, Match Keys, and the Attribution Windows Nobody Sets
TikTok delivery learns from what the Events API sends, scores it as Event Match Quality, and attributes it through click, engaged-view, and view windows most accounts never chose. The pixel plus Events API stack, the payload that lifts match quality, deduplication that actually works, the windows to set per ad group, and what changed in 2025 and 2026.
The UTM Naming Convention Standard for Performance Marketing Teams (2026)
One tagging standard that works across Google Ads, Meta, TikTok, LinkedIn, email, and creators, lands correctly in GA4's channel groups, survives Apple's click-id stripping, and can be enforced by an agent. Field specs, per-platform templates, and the rules that keep it clean.
Identity stitching
The part of your analytics stack that's almost certainly broken — five layers, three bugs we caught in review, and the SQL to copy.
Identity lock
Cameo, character lock, or text-only — the decision that silently kills AI video ads, and how to pick right.
The wrong customer for AI video
HeyGen, Synthesia, Higgsfield, Runway, Sora — what they're actually optimized for, and why advertising is still waiting for the right tool.